Removing T-Cell Epitopes with Computational Protein Design

Removing T-cell epitopes with computational protein design

Chris Kinga,1, Esteban N. Garzab, Ronit Mazorc, Jonathan L. Linehand, Ira Pastanc, Marion Pepperb, and David Bakera

aInstitute for Protein Design, Department of Biochemistry and bDepartment of Immunology, University of Washington, Seattle, WA 98195; and cNational Cancer Institute and dNational Institute of Allergy and Infectious Disease, National Institutes of Health, Bethesda, MD 20892

Edited by Barry Honig, Howard Hughes Medical Institute, Columbia University, New York, NY, and approved April 24, 2014 (received for review December 4, 2013)

Immune responses can make protein therapeutics ineffective or epitope data, MHC epitope prediction tools, and host genomic even dangerous. We describe a general computational protein design data to enable the design of proteins with reduced immunoge- method for reducing immunogenicity by eliminating known and nicity while retaining function and stability. Our approach goes predicted T-cell epitopes and maximizing the content of human beyond PROPRED by implementing a more accurate machine peptide sequences without disrupting protein structure and func- learning-based epitope prediction method for 28 different H-2, tion. We show that the method recapitulates previous experi- HLA-DR, and HLA-DQ alleles, restricts design to select 15mer mental results on immunogenicity reduction, and we use it to epitope regions, and uses a greedy stepwise protein design al- disrupt T-cell epitopes in GFP and Pseudomonas exotoxin A with- gorithm (10) to eliminate the most immunogenic epitopes with out disrupting function. as few mutations as possible, avoiding disruptive core mutations likely to destabilize the protein. We compare the performance of deimmunization | machine learning | biotherapeutics | Rosetta | our epitope predictor with PROPRED and another leading epi- immunotoxin tope prediction method for 13 different human and mouse MHC alleles, show the effectiveness and generality of the method with in silico tests on previously characterized deimmunized protein tar- mmunogenicity is a major problem in the development of pro- gets, and show experimentally for GFP in mice and Pseudomonas Itein therapeutics. Repeated administration of a protein thera- exotoxin A (PE38) in humans that the method eliminates T-cell peutic can lead to B-cell activation and production of antibodies, epitopes without disrupting function. rendering the therapeutic clinically ineffective or cross-reacting with host proteins (1). Affinity maturation of antibody-producing Results memory B cells is initiated by T-cell recognition of peptide epitopes Overview of Computational Method. For a given target protein and displayed on major histocompatibility complex class II (MHCII) set of host MHC alleles, potential T-cell epitopes are first identified proteins on the surface of mature antigen-presenting cells. Immu- using a support vector machine (SVM). These regions are then nogenicity may be reduced by eliminating known T-cell epitopes optimized to eliminate the T-cell epitopes while retaining structure from the protein sequence and/or increasing the prevalence of and function using an extension of the Rosetta all-atom protein sequences already found in the host genome to which T cells would design methodology with modifications to both the energy function already be tolerant, an approach that has met with substantial used in the design calculations and the optimization procedure. clinical success in the humanization of recombinant antibodies (2). The energy function used in the sequence optimization is However, unlike antibodies, which have been extensively charac- supplemented with two terms that incorporate immunologically terized, the mutational tolerance of most proteins is generally not relevant data. The first term calculates predicted epitope content using SVMs trained with experimentally determined peptide– known, and hence, the extension of this approach to proteins of MHC binding data. Scores from SVMs for each MHC allele in arbitrary structure and function remains a major challenge. Dei- each 15mer sequence frame are averaged and then summed over mmunization efforts have relied, for the most part, on experimental characterization of a large number of point mutants Significance followed by a combination of individual mutations (3, 4). To reduce or eliminate immunogenicity, it would be desirable to have a method that eliminates MHCII-binding epitopes and Proteins represent the fastest-growing class of pharmaceuticals increases host sequence content without disrupting interactions es- for a diverse range of clinical applications. Computational pro- sential for proper folding and function. The peptide-binding rep- tein design has the potential to create a novel class of ther- ertoire of many MHCII alleles has been extensively characterized apeutics with tunable biophysical properties. However, the (5), and a number of methods has been developed for predicting the immune system reacts to T-cell epitope sequences in non- affinity of novel peptides for a given MHCII (6). Coupling of epi- human proteins, leading to neutralization and elimination by tope prediction methods with methods for predicting the structural the immune system. Here, we combine machine learning with and functional consequences of mutations offers the possibility structure-based protein design to identify and redesign T-cell of reducing the immunogenicity of a target protein without epitopes without disrupting function of the target protein. We disrupting structure and function. Epitope prediction methods, test the method experimentally, removing T-cell epitopes from homolog substitution matrices, and structural stability calcu- GFP and Pseudomonas exotoxin A while maintaining function. lations have been combined to predict optimal epitope-eliminating mutations (7, 8). Epitope prediction methods have been inte- Author contributions: C.K., E.N.G., R.M., I.P., M.P., and D.B. designed research; C.K., E.N.G., grated with structure-based protein design (9) by combining the and R.M. performed research; C.K., E.N.G., R.M., and J.L.L. contributed new reagents/ COMPUTER SCIENCES 9mer epitope PROPRED matrices with protein design of all resi- analytic tools; C.K., E.N.G., R.M., I.P., M.P., and D.B. analyzed data; and C.K. wrote dues in a flexible backbone method that allows substantial redesign the paper. of protein cores. The combined method was able to eliminate epi- The authors declare no conflict of interest. tope-like sequences while maintaining native-like values for a num- This article is a PNAS Direct Submission. ber of predicted protein stability metrics, but folding, function, Freely available online through the PNAS open access option. and immunogenicity were not evaluated experimentally. 1To whom correspondence should be addressed. E-mail: [email protected]. Here, we describe the integration of the Rosetta computa- This article contains supporting information online at www.pnas.org/lookup/suppl/doi:10. MEDICAL SCIENCES tional protein design method with experimental immunogenic 1073/pnas.1321126111/-/DCSupplemental.

www.pnas.org/cgi/doi/10.1073/pnas.1321126111 PNAS | June 10, 2014 | vol. 111 | no. 23 | 8577–8582 Downloaded by guest on September 26, 2021 PROPRED achieved significantly lower prediction accuracy (AUC = 0.710). Because more T-cell epitopes have been characterized for some alleles in the test set, combining all testing data weights performance analysis with alleles with more data points. If we average AUCs with equal weight over the shared DRB allele set, again, NetMHCII performs best (AUC = 0.792), with Rosetta slightly lower (AUC = 0.785) and PROPRED lower still (AUC = 0.771) (Fig. 1B).

Large-Scale Benchmarking and Calibration of Design Method. We first tested the ability of the method to eliminate putative human T-cell epitopes from eight proteins from pathogenic organisms that contain known MHC-binding epitopes. The computational design protocol was used to simultaneously eliminate all predicted Fig. 1. Performance of Rosetta SVM T-cell epitope prediction. (A) ROC curve epitopes for eight representative human DRB1 alleles (SI Meth- true-positive rate vs. false-positive rate for all testing data and comparison ods), collectively covering almost 95% of the human population with current methods. Total AUC is listed for each method. (B) Predictive (15). To evaluate the tradeoff between epitope removal and pro- performance over each allele test set. x Axis, Rosetta AUC; y axis, other tein stability, multiple design simulations were carried out, with an method AUC. Points below the 1:1 dotted line indicate where Rosetta performs better than other methods. increasing weight on the SVM-based epitope scoring term. In- creasing the weight on this term decreases the number of MHCII predicted epitopes and increases the Rosetta energy (Fig. 2A). Because amino acid substitutions predicted to disrupt pep- each frame. The second term uses known host genome 9mer – data and known epitope data, rewarding each host 9mer in pro- tide MHC binding might destabilize the overall protein, a bal- portion to its frequency of occurrence in the host genome and ance between the stability of the protein and disruption of possible penalizing known epitopes. Both deimmunization scores favor MHC binding must be sought. A weight on the epitope scoring negatively charged residues on the surface of the protein; hence, term of 2.0 eliminates 79% of the predicted epitopes and 84% of we also introduce a net charge constraint into the total objective the known epitopes without increasing Rosetta energy above that function, penalizing deviations from the input protein formal of the native protein (Table S1). Because sequence changes are charge. By weighting this term appropriately against the deimmu- permitted only at critical predicted epitope positions, the num- nization terms, negatively charged residues may be placed at ber of mutations is minimized, thus allowing for substantial re- critical epitope-disrupting surface positions while compensatory duction in predicted immunogenicity while maintaining average positively charged residues are introduced at other positions. sequence identity at 66%. Similar calculations were carried out Sequence optimization is carried out using a protocol that fo- varying the weight of the term favoring 9mer sequences found in cuses on solutions that reduce the energy with a relatively small the human genome (Fig. 2B). As the weight increases, the average number of mutations and allows use of computationally expensive number of human genome 9mers over the designed regions in- objective functions in design calculations that otherwise might be creases, and the Rosetta energy becomes more unfavorable. At a unacceptably slow. The energies of all point mutants at each de- weight of 3.5, human 9mer sequences increase from 0% to 4.3% sign position are first computed and then sorted. At each position, of redesigned epitopes, whereas the average Rosetta energy all point mutants within a certain threshold of the lowest energy increases only 10% over baseline (Table S2). These weights were mutant are saved for subsequent combination. These lowest en- used for the remainder of this work unless otherwise noted ergy point mutations are then combined using a greedy stepwise, in Methods. steepest descent heuristic (Methods), allowing for structural re- laxation at each step, until mutations have been attempted at all Recapitulation of Previous Immunogenicity Reduction Data. We next design positions. Multiple diverse near-optimal designs can be attempted to recapitulate the results of previous experimentally generated in parallel by stochastically accepting the placement of validated protein deimmunization efforts where immunogenicity near-optimal mutations during the combination process. was reduced without disrupting biological function. Cantor et al. (16) sought to remove T-cell epitopes from Escherichia coli T-Cell Epitope Prediction. We initially trained our SVMs on experi- L-asparaginase II, an enzyme approved for treatment of acute mentally measured MHC binding affinities for 26 allelic variants lymphoblastic leukemia, while maintaining native-like enzymatic (11). For each MHC allele type represented in the training set, we activity and protein stability. Cantor et al. (16) first used a neutral collected all epitope sequences from the Immune Epitope Database drift selection scheme to identify allowable mutations in each of (IEDB) known to elicit T-cell activation (5). We restricted addi- three predicted HLA-DRB1*04:01 epitopes and then combined tional analysis to alleles with at least 25 known T-cell epitopes in the IEDB. In the absence of sufficient data on nonbinding peptides for each allele, we assumed that most 15mers in the host genome would not constitute strong MHC binders and generated negative datasets by randomly choosing 1,000 15mers from the human genome (12). We compared our SVM-based epitope predictions with the predictions of previous methods using a subset of T-cell epitopes withheld from initial training. Sensitivity and specificity were evaluated over all alleles for our method, NetMHCII-v2.2 (13), and PROPRED (14), and predictive performance was evaluated by calculating the area under the receiving operator characteristic (ROC) curve (AUC) for each allele type both independently and over the entire set. Because PROPRED contains only matrices for human leukocyte antigen DR beta (HLA-DRB) alleles, we combined testing data for the eight DRB alleles covered by all three Fig. 2. Tradeoffs between Rosetta energy and extent of deimmunization. methods and generated a standard ROC plot (Fig. 1A). The Rosetta energy (□) of redesigned proteins increases, whereas (A) epitope highest AUC was achieved by NetMHCII (AUC = 0.759). Rosetta content decreases and (B) human 9mer count increases (♦) as the weights on performed comparably but slightly worse (AUC = 0.752), whereas the associated score terms are increased.

8578 | www.pnas.org/cgi/doi/10.1073/pnas.1321126111 King et al. Downloaded by guest on September 26, 2021 mutations in each epitope to produce a fully functional enzyme concern about the immunogenicity of cells expressing GFP remains with reduced immunogenicity in vivo. We applied our protocol to (19).WesoughttoredesignsfGFPto eliminate T-cell epitopes the biologically active tetrameric state of E. coli L-asparaginase II, without disrupting fluorescence. To do so, we targeted the top constraining design to the residue positions chosen for randomi- four predicted H-2-IAb epitopes in the sfGFP sequence (Fig. zation in the previous study, and compared our predictions against 3A). None of these epitopes were present in our known epitope the mutations identified in the selection screen and the predicted database, although epitope 84 had been previously identified immunogenicity of the resultant design sequence. After Rosetta as immunodominant in WT GFP. The design algorithm is nearly redesign for all 24 epitope residues, two of five mutations assumed deterministic; to generate multiple candidates for testing, we residue identities found in the sequences of experimentally isolated stochastically sampled alternative sequences by random inclusion activity-preserving variants, with one mutation occurring at a posi- of locally near-optimal mutations at each design position. Eight tion not tested in the above study. In addition, two of three rede- designs were chosen for testing based on sequence diversity and signed epitopes have predicted HLA-DRB1*04:01 binding affinity predicted stability (Table 2 and SI Methods, sfGFP Deimmunization lower than the peptides tested in the previous study (Table 1). Design Sequences Alignment). sfGFP and all eight designs were Tangri et al. (4) attempted the deimmunization of erythro- expressed in E. coli and purified as soluble protein. To determine poietin (Epo), a growth hormone used in treatment of anemia whether fluorescence was affected by the design mutations, and myelodysplasia. Tangri et al. (4) screened a set of Epo-derived emission and absorbance spectra were obtained for sfGFP and peptides for binding to a set of 15 different HLA alleles, targeted all eight deimmunized proteins. All eight designs showed fluores- two potential epitopes for deimmunization by screening a small cence absorbance and emission spectra comparable with sfGFP, set of point mutants in each epitope for reduction of T-cell acti- with fluorescence excitation peaks at 485 nm (Fig. 4A and Fig. S1). vation, combined point mutants, and screened for both immuno- We then investigated whether existing epitopes were correctly genicity and biological activity. We applied our deimmunization identified and removed and whether new epitopes were not in- design protocol to this protein using the structure of Epo bound to troduced by the design mutations. sfGFP and the deimmunized its native receptor and targeting the same epitope positions ex- variant (sfGFP.di.v3.2) were chosen for immunological testing. perimentally explored by Tangri et al. (4). Because the HLA allele This variant was selected as the most aggressive design, because set of the human donors in the study is unknown, we designed it had the highest Rosetta energy but still maintained function. simultaneously against all eight alleles in the HLA-DRB1 allele GFP:I-Ab tetramer reagents were generated using both native set. In the first epitope region, Rosetta introduces only one sub- and design peptide sequences for three of the predicted epitope stitution (the same mutation found to be optimal in ref. 4). Rosetta regions. For both constructs, five mice were injected with the similarly recovers the point mutation position in the second epitope protein in complete Freund’s adjuvant (CFA). After 6 d, spleens (Table 1) but substitutes an alanine to preserve the protein’snet from all 10 mice were stained with a multicombinatorial panel of charge and makes three additional mutations to further disrupt GFP:I-Ab tetramers corresponding to both the native and design MHC binding. In both cases, the Rosetta design minimizes the sequences of all three predicted epitope regions (Table 2), and number of predicted MHC-binding peptides while minimizing tetramer-positive cells were magnetically bead-enriched. For mice the energy of the protein in the epitope regions. challenged with WT sfGFP, flow cytometry confirmed epitope 84 as the immunodominant epitope, with epitope 223/224 recognized Epitope Removal in Superfolder GFP. As an experimental proof of by a smaller number of T cells (Fig. 4B). For mice challenged with concept, we chose the fluorescent reporter protein superfolder the deimmunized protein, all three tetramers corresponding to the GFP (sfGFP) (17). GFP is used to identify and track genetically three redesigned epitope regions failed to isolate T cells above modified stem cells in vivo for numerous applications (18) but background levels, except for one mouse that responded weakly to

Table 1. Recapitulation of previous E. coli L-asparaginase II deimmunization efforts (16) against one HLA allele and Epo (4) against eight HLA alleles

Sequence Predicted IC50 (nM) Rosetta energy

EcAII: Epitope 1 (115–123) (rank 6) Native MRPSTSMSA 194.3 −12.0 Cantor et al. (16) VRPPTRMSP 339.9 77.4 Rosetta MRPQTFMSA 87.2 −9.4 EcAII: Epitope 2 (216–224) (rank 11) Native IVYNYANAS 217.3 −15.4 Cantor et al. (16) VVYGYANAS 195.8 −13.8 Rosetta IVYNYSNAM 197.1 −11.2 EcAII: Epitope 3 (304–312) (rank 3) Native VLLQLALTQ 135.3 −17.4 Cantor et al. (16) VLLTLALTN 122.4 −11.3 Rosetta VLLQLALWQ 193.1 −13.4 Epo: Epitope 1 (101–115) (rank 7) Native GLRSLTTLLRALGAQ 7.7 −20.0 Tangri et al. (4) GLRSLTDLLRALGAQ 12.1 −20.3 Rosetta GLRSLTDLLRALGAQ 12.1 −20.3 Epo: Epitope 2 (136–150) (rank 1) Native DTFRKLFRVYSNFLR 5.0 −23.5 Tangri et al. (4) DTFRKLFRVYDNFLR 24.0 −19.9 Rosetta DTFRKEFFDYANFLR 70.2 −13.7

MHC IC50 values are predicted by Rosetta SVM, and Rosetta energies are the sum of total residue energies over the epitope region. IC50 values for Epo are listed as the lowest predicted across the allele set. Epitope ranks as a function of predicted immunogenicity are listed next to the residue ranges. Mutations are highlighted in MEDICAL SCIENCES COMPUTER SCIENCES bold, and known activity-preserving mutations are italicized. EcAII, E. coli L-asparaginase II.

King et al. PNAS | June 10, 2014 | vol. 111 | no. 23 | 8579 Downloaded by guest on September 26, 2021 the WT and mutant epitope peptides. Both mutants caused a significant reduction in T-cell response for both epitope peptides in all three samples (P > 0.01 in Student t test) (Fig. 5B). Discussion We have developed a computational protein design method that incorporates host genome information and MHC-binding prediction tools to reduce the immunogenicity of arbitrary protein targets. The method removes MHC epitopes and increases human sequence content while maintaining protein stability and interactions with binding partners. Mutations predicted by the method partially recapitulate the mutations of previous suc- cessful deimmunization efforts. The effectiveness of the method was verified experimentally by successfully predicting and eliminating an immunodominant T-cell epitope from sfGFP and eliminating a known T-cell epitope from PE38. We redesigned these proteins to mitigate T-cell immune responses in both mice and humans, respectively, showing that the method presented is generally applicable to humans or other species for which sufficient MHC data exist. Fig. 3. Rosetta design model for sfGFP deimmunization. (A) Published coor- Because of the possibility of disrupting folding and function of dinates of sfGFP crystal structure. Both known and predicted epitopes were the target protein, most deimmunization efforts rely on experi- targeted for design. Epitope indices from Table 2 are labeled in circles. (B) mental testing of a limited number of point mutants followed by Close-up view of immunodominant epitope. (C) Rosetta deimmunization de- conservative attempts to combine a small subset of these mutations sign of B. Cyan, design mutations; green, sfGFP; magenta, predicted epitopes. into a functional product. Although varied in their approaches, proprietary methods used by companies, such as EpiVax (25) and Antitope (26), typically combine matrix-based epitope scoring, the mutant epitope 223/224 (Fig. 4C). These two experiments experimental characterization of MHC binding or T-cell epitope confirmed that the design mutations effectively eliminated or mapping, stepwise mutation of antigenic amino acids, and in- greatly reduced the antigenicity of WT T-cell epitopes without troduction of tolerizing epitopes where possible. Using structure- creating new immunodominant epitopes or disrupting fluorescence. based design simulations, we introduced nine mutations into sfGFP simultaneously without disrupting function. One rede- Epitope Removal in Domain III of PE38. We next sought to remove signed epitope is largely buried in the core of the protein (Fig. 3 B T-cell epitopes from the toxin domain of the cancer therapeutic and C), requiring the selection of mutations that simultaneously HA22, a recombinant immunotoxin containing a 38-kDa frag- eliminates MHC-binding propensity without disrupting the fold- ment of PE38 (20), while maintaining cytotoxic activity. HA22 ing free energy of the protein. Eliminating such epitopes would has been used successfully to treat refractory hairy cell leukemia be difficult using the above proprietary approaches, because in a recent phase 1 clinical trial (21), and it has produced complete multiple substitutions may be required to compensate for remission in several patients with acute lymphoblastic leukemia packing defects caused by mutations in the epitope region, (22). However, HA22 has shown limited effectiveness in treating and nonconservative mutations near the protein-binding in- patients who are not immune-compromised, because the pres- terface of exotoxin A would not have been predicted without ence of the bacterial PE38 moiety leads to host immune response a structure-based design calculation. and the production of neutralizing antidrug antibodies (23). To This work has focused only on eliminating the most immu- address this issue, we targeted three epitope regions in PE38 noreactive epitopes for a given set of MHC alleles. Immuno- previously identified in humans (24) and designed five mutations logical testing of a larger number of human patients would be predicted to eliminate binding to a diverse set of 14 human HLA required to fully cover the breadth of HLA allotype diversity, alleles while maintaining favorable interactions with the toxin and there may exist highly conserved T-cell reactive amino acids substrate, eukaryotic elongation factor 2 (Fig. S2). The five in protein sequences for which no immune silencing mutations mutants were expressed and purified for subsequent testing of are possible, necessitating the prediction and removal of discon- cytotoxicity in two Burkitt lymphoma cell lines. Two mutants in tinuous B-cell epitopes from the protein surface. Methods exist for the region 466–480, A476D and A476D+D474Y, had cytotoxic prediction of B-cell epitopes but suffer from lower predictive activity reduced by ∼80%, but three mutants in the region 547– power because of the difficulty of obtaining sufficient structural 564 displayed equal or greater cytotoxicity than the WT toxin data for the entire repertoire of discontinuous 3D epitopes (27). (Fig. 5A). The two most active mutants (L552N and L552E) Nevertheless, B-cell epitope removal methods have proven suc- were chosen for additional characterization of immunogenicity. cessful for a number of clinical targets (28, 29). Such methods could Peripheral blood mononuclear cells (PBMCs) derived from two be incorporated into a comprehensive deimmunization pipeline, patients and one naïve donor were stimulated with mutant an- and additional testing of systemic antibody response would be re- tigen, and IL-2 response was measured after restimulation with quired to show complete immune evasion.

Table 2. sfGFP epitopes targeted for redesign sfGFP sfGFP.di.v3.2

Index Native sequence Predicted IC50 (nM) Rosetta energy Design sequence Predicted IC50 (nM) Rosetta energy

8 FTGVVPILV 548 −12.9 FKGRVPIQV998−10.6 84 FKSAMPEGY 784 −8.4 MKSAMPDGY 4,465 −8.9 223 FVTAAGITH 542 −8.4 FVRAAGIQE 3,312 −7.9 224 VTAAGITHG 954 −7.15 VRAAGIQEE 2,354 −6.5

Mutations are highlighted in bold.

8580 | www.pnas.org/cgi/doi/10.1073/pnas.1321126111 King et al. Downloaded by guest on September 26, 2021 models in the user-defined allele list, and summing the contributions from each overlapping sequence frame.

Rewarding 9mer Sequences That Occur in the Host Genome and Penalizing Known Epitopes. Host 9mer database construction. Protein translations of Ensembl gene predictions for Homo sapiens and Mus musculus genomes were downloaded from ftp://ftp.ensembl.org/pub/current_fasta/ on February 28, 2012. The number of occurrences of every unique contiguous 9mer peptide found in all hy- pothetical translation products was first calculated. Each 9mer was assigned a score that rewards common sequences; 9mers that occur between 1 and 10 times were assigned a score of −log(n) − 1, where n is the total count of the 9mer, and 9mers that occur more than 10 times were given a constant score of −2.0 to prevent domination of scoring by widespread repeat sequences. Thus, each unique sequence was given a score in the range [−2, −1]. Known epitope database construction. All epitope sequences known to elicit T-cell activation through the MHCII pathway in either humans or mice were downloaded from the IEDB on February 28, 2012; 9mer core sequences were predicted as the epitope subsequence with the highest predicted MHC binding affinity as scored by Rosetta SVMs. All 9mer epitope sequences were assigned a constant score of 10.0. 9mer Database scoring. 9mer Subsequences with associated scores (genomic 9mers and known epitope sequences) are loaded from a user-supplied table at runtime and stored in a hash table for quick look-up during design. The Fig. 4. Redesign of sfGFP reduces T-cell reactivity without disrupting fluo- total score is the sum of the scores for each 9mer subsequence. rescence. (A) Deimmunized sfGFP excitation and emission spectra in arbitrary units (AU). Excitation spectrum measured at 510-nm emission. Emission Rosetta Design Calculations. Structure preprocessing and simulation parameters. spectra measured at 488-nm excitation. (B) Flow cytometry analysis of tet- + + + + Before design calculations, all protein structures were subject to multiple ramer-enriched populations of CD3 CD4 CD44 GFP:I-Ab lymphocytes + cycles of backbone minimization and rotamer optimization with position labeled with phycoerythrin (PE) and/or A-phycocyanin (APC). Total CD44 + restraints on side-chain heavy atoms, allowing for small structural changes to CD4 tetramer-positive cells for each of six GFP:I-Ab tetramers in mice im- bring the structure to the local energy function minimum. For all design munized with WT sfGFP. Immunization with the native sfGFP leads to the + calculations, talaris2013 Rosetta score weights were used; native residues expansion and activation of CD4 T cells responding to epitopes 82–96. (C) + + were given a constant energy bonus of −0.2 Rosetta energy units (REU), Total CD44 CD4 tetramer-positive cells for each of six GFP:I-Ab tetramers in nonnative residues were given a penalty of 0.2 REU, and nonnative cysteine mice immunized with the designed sfGFP 3.2. Mice immunized with the and histidine residues were disallowed in design. designed sfGFP 3.2 no longer respond to the native sfGFP 82–96 epitopes or Rosetta greedy optimization design. All design simulations were implemented the designed epitopes 82–96 in sfGFP 3.2. MUT, mutant. using Rosetta Scripts (32). Greedy sequence design and rotamer optimization were carried out using the Rosetta greedy descent optimization algorithm as previously described (11). Details are provided in SI Methods. Briefly, every Here, we have focused on application to protein therapeutics amino acid point mutant is sampled independently, and the total energy is delivered extracellularly, whereMHCIImediatestheprimaryim- stored. Optimal mutations are sorted by energy before combinatorial opti- munological pathway. The method is readily extensible to both mization is attempted in a rank-ordered, steepest descent fashion. Multiple MHC class I- and MHCII-based deimmunization, and it is, thus, diverse solutions can be generated by stochastically attempting combination applicable to therapeutics expressed intracellularly, such as gene of near-optimal substitutions at each position. therapy products. Because epitope scores are averaged over all Deimmunization design simulations. Details of each design calculation are MHC allele SVMs, degenerate binding epitopes are penalized provided in SI Methods. Briefly, all simulations followed a similar workflow more strongly and thus, are the first to be targeted in our greedy as follows. Crystal structures were downloaded from the Protein Data Bank design algorithm. This scoring scheme is critical, because mounting evidence points to the correlation between the number of host MHC alleles that a given epitope binds and its propensity to initiate an immune reaction in vivo (4, 30). Our epitope prediction method was trained on large-scale peptide-binding affinity data, although predictions can be made using other sources. When experimental binding constants are not available, MHC–peptide structure simulations have shown promise in calculating accurate sequence specificities based on MHC–peptide energetics (31). As improvements in energy functions lead to improvement in the prediction of the effects of mutations on stability and function and high-throughput experimental MHC–peptide-binding data become increasingly available, computational protein design will play an increasingly prominent role in development of next generation protein therapeutics. Methods Penalizing Predicted Epitopes. Epitope SVM construction. A detailed description of SVM construction is provided in SI Methods. Briefly, one SVM model was trained for each MHC using publicly available peptide–MHC binding constants for 29 human and mouse MHC alleles; 15mer peptide sequences were first aligned and then encoded into numerical feature vectors as described in SI Methods. After encoded, SVM regression models were trained using

libSVM to recapitulate peptide–MHC binding IC50 values by minimizing Fig. 5. Redesign of exotoxin A reduces T-cell reactivity without loss of mean-squared error in fivefold cross-validation tests. function. (A) Relative cytotoxicity for the original HA22-LR toxin and three Epitope SVM scoring. The total epitope prediction score of a protein during computationally designed variants in two cell types. (B) ELISpot IL-2 response Rosetta design is calculated by sliding a 15-residue window across the protein was measured for PBMCs derived from two patients and one naïve donor sequence, calculating the average SVM binding score over all allele SVM after restimulation with two WT peptides and four mutant peptides. MEDICAL SCIENCES COMPUTER SCIENCES

King et al. PNAS | June 10, 2014 | vol. 111 | no. 23 | 8581 Downloaded by guest on September 26, 2021 and preminimized with Rosetta as described above. Designable residues were Cell enrichment and flow cytometry. All antibodies were from eBioscience unless selected on the basis of average predicted MHC binding affinity; all residues in noted. Single-cell suspensions of spleens and lymph nodes were stained for 1 h at any epitope frame that score above a certain threshold are selected for design. room temperature with eGFP:I-Ab allophycocyanin tetramer. Samples were For comparison with previous experimental works, residues were selected so as then enriched for bead-bound cells on magnetized columns, and a portion was to provide a meaningful comparison with the published data. Deimmuniza- removed for counting as described (33). For identification of surface pheno- tion design was then carried out as described above. For design targets un- type, the rest of the sample underwent surface staining on ice with a mixture dergoing experimental characterization, multiple design sequences were of antibodies specific for B220 (RA3-6B2), CD11b (MI-70), CD11c (N418), CD44 generated by randomly sampling near-optimal mutations at each design po- (IM7; BD), CD4 (RM4-5; BD), CD3 (145-2C11), and CD8 (5H10; BioLegend), each sition and combining these mutations in a stochastic manner. conjugated with a different fluorochrome. Cells were then analyzed on a Canto (BD) flow cytometer. Data were analyzed with FlowJo software (TreeStar). b sfGFP Activity and Immunogenicity Assays. Identification of eGFP82–96:I-A epitope. Fluorescence spectra. sfGFP samples were diluted to a uniform concentration of To begin to narrow down a region in the EGFP protein containing a CD4+ 9.8 μM in PBS, and fluorescence spectra were measured on a SpectraMax T-cell epitope in C57BL/6 mice, the EGFP nucleotide sequence (Clontech) was plate reader. Excitation was measured from 448 to 500 nm (2-nm intervals) parsed into five equally sized fragments, and each fragment was inserted into at 510-nm emission. Emission was measured from 498 to 550 nm (2-nm the TOPO cloning site of the pTrcHis2-TOPO vector containing an isopropyl intervals) at 488-nm excitation. Fluorescence spectra were normalized by beta-D-1 thiogalactopyranoside-inducible promoter and a 6× His C-terminal subtracting the signal obtained from pure PBS buffer. epitope tag (Invitrogen). EGFP protein fragments were expressed in Rosetta 2 competent E. coli (EMD Millipore). Bacteria were lysed with BugBuster protein Exotoxin A activity and immunogenicity assays. Construction, expression, purification, extraction reagent (Novagen) and sonicated, and EGFP fragments were purified and cytotoxic activity of recombinant immunotoxin. The mutations L552E, L552N, with His-Bind resin columns (Novagen). Next, individual mice were immunized and P559E were introduced into a plasmid expressing HA22 VH-PE38 using PCR s.c. with 25 μg purified whole EGFP protein (Biovision, Inc.) in CFA (Sigma). After overlap extension. The mutant RIT were purified as previously described (34). + + 10 d, draining lymph node cells were negatively selected for CD4 Tcells Cytotoxicity assays were performed on CD22 human Burkitt lymphoma cell (Miltenyi Biotech), and an anti–IFN-γ ELISPOT assay was done by interrogating lines (CA46 and Raji). The assay was performed as previously described (24). with the above purified EGFP protein fragments (antibodies were from eBio- In vitro expansion and ELISpot. PBMCs from two patients who were previously science, and 96-well Multiscreen filter plates were from Millipore). On identi- treated with PE38 RIT and one naïve donor were obtained after informed con- + fying the EGFP protein fragment that gave the best IFN-γ–producing CD4 T-cell sent. The PBMCs were stimulated with parent RIT, HA22-L552E, or HA22-L552N

response, a 15mer overlapping peptide library (each offset by 2 aa) was con- andculturedin37°Cwith5%(vol/vol)CO2 for 14 d. IL-2 (10 U/mL; Millipore) was structed (Mimotopes). This library was then used for interrogation in another added every 3 d. On day 14, cells were harvested and restimulated with either ELISPOT assay, as described above, to narrow down the 15mer epitope to WT peptides 93 and 94 (GPEEEGGRLETILGW and EEGGRLETILGWPLA) or amino acids 82–96. mutant peptides (GPEEEGGREETILGW, EEGGREETILGWPLA, GPEEEGGRNETILGW, Immunizations. C57BL/6 mice were injected s.c. at the base of the tail with and EEGGRNETILGWPLA). IL-2 ELISpot was used to detect T-cell activation either 50 μg sfGFP or 50 μg sfGFP.di.v3.2 emulsified in 50 μL CFA (Sigma- accordingtothemanufacturer’s instructions. Aldrich). C57BL/6 mice were injected i.p. with 100 μg sfGFP or 100 μg sfGFP. Rosetta command line demonstration. Command line examples, input files, and di.v3.2 in aluminum hydroxide adjuvant (Brenntag). instructions for running all protein design simulations are included in a freely Tetramer production. Biotin-labeled soluble I-Ab molecules containing EGFP available archived demonstration at https://zenodo.org/record/8436. peptide (FKSAMPEGY) covalently attached to the I-Ab β-chain were produced in Drosophila melanogaster S2 cells and then purified and made into ACKNOWLEDGMENTS. This research was supported by the Defense Threat tetramers with streptavidin-phycoerythrin or streptavidin-allophycocyanin Reduction Agency and the Intramural Research Program of the National (Prozyme) as described (33). Institutes of Health, the National Cancer Institute, Center for Cancer Research.

1. Zubler RH (2001) Naive and Memory B Cells in T-Cell-Dependent and T-Independent 19. Persons DA, et al. (1998) Use of the green fluorescent protein as a marker to identify Responses. Springer Seminars in Immunopathology (Springer, Berlin), pp 405–419. and track genetically modified hematopoietic cells. Nat Med 4(10):1201–1205. 2. Goldenberg MM (1999) Trastuzumab, a recombinant DNA-derived humanized mono- 20. FitzGerald DJ, Wayne AS, Kreitman RJ, Pastan I (2011) Treatment of hematologic clonal antibody, a novel agent for the treatment of metastatic breast cancer. Clin Ther malignancies with immunotoxins and antibody-drug conjugates. Cancer Res 71(20): 21(2):309–318. 6300–6309. 3. Harding FA, et al. (2005) A beta-lactamase with reduced immunogenicity for the 21. Kreitman RJ, et al. (2012) Phase I trial of anti-CD22 recombinant immunotoxin mox- targeted delivery of chemotherapeutics using antibody-directed enzyme prodrug etumomab pasudotox (CAT-8015 or HA22) in patients with hairy cell leukemia. J Clin therapy. Mol Cancer Ther 4(11):1791–1800. Oncol 30(15):1822–1828. 4. Tangri S, et al. (2005) Rationally engineered therapeutic proteins with reduced im- 22. Wayne AS, et al. (2011) A novel anti-CD22 immunotoxin, moxetumomab pasudotox: munogenicity. J Immunol 174(6):3187–3196. Phase I study in pediatric acute lymphoblastic leukemia (ALL). ASH Annual Meeting 5. Vita R, et al. (2010) The immune epitope database 2.0. Nucleic Acids Res 38(Database Abstracts 118(21):248. issue):D854–D862. 23. Hassan R, et al. (2007) Phase I study of SS1P, a recombinant anti-mesothelin im- 6. Nielsen M, Lund O, Buus S, Lundegaard C (2010) MHC class II epitope predictive al- munotoxin given as a bolus I.V. infusion to patients with mesothelin-expressing gorithms. Immunology 130(3):319–328. mesothelioma, ovarian, and pancreatic cancers. Clin Cancer Res 13(17):5144–5149. 7. Parker AS, Griswold KE, Bailey-Kellogg C (2011) Optimization of therapeutic proteins 24. Mazor R, et al. (2012) Identification and elimination of an immunodominant T-cell to delete T-cell epitopes while maintaining beneficial residue interactions. J Bioinform epitope in recombinant immunotoxins based on Pseudomonas exotoxin A. Proc Natl Comput Biol 9(2):207–229. Acad Sci USA 109(51):E3597–E3603. 8. Parker AS, Zheng W, Griswold KE, Bailey-Kellogg C (2010) Optimization algorithms 25. De Groot AS, Terry F, Cousens L, Martin W (2013) Beyond humanization and de- for functional deimmunization of therapeutic proteins. BMC Bioinformatics 11(1):180. immunization: Tolerization as a method for reducing the immunogenicity of bio- 9. Choi Y, Griswold KE, Bailey-Kellogg C (2013) Structure-based redesign of proteins for logics. Expert Rev Clin Pharmacol 6(6):651–662. minimal T-cell epitope content. J Comput Chem 34(10):879–891. 26. Baker M, Carr F (2010) Pre-clinical considerations in the assessment of immunoge- 10. Nivón LG, Bjelic S, King C, Baker D (2014) Automating human intuition for protein nicity for protein therapeutics. Curr Drug Saf 5(4):308–313. design. Proteins 82(5):858–866. 27. Greenbaum JA, et al. (2007) Towards a consensus on datasets and evaluation metrics 11. Wang P, et al. (2008) Immune epitope database: MHC-II binding dataset. Available at for developing B-cell epitope prediction tools. J Mol Recognit 20(2):75–82. http://tools.iedb.org/mhcii/download/. Accessed February 14, 2013. 28. Nagata S, Pastan I (2009) Removal of B cell epitopes as a practical approach for re- 12. Flicek P, et al. (2012) Ensembl 2012. Nucleic Acids Res 40(Database issue):D84–D90. ducing the immunogenicity of foreign protein-based therapeutics. Adv Drug Deliv 13. Nielsen M, Justesen S, Lund O, Lundegaard C, Buus S (2010) NetMHCIIpan-2.0— Rev 61(11):977–985. Improved pan-specific HLA-DR predictions using a novel concurrent alignment and 29. Onda M, et al. (2008) An immunotoxin with greatly reduced immunogenicity by weight optimization training procedure. Immunome Res 6(1):9. identification and removal of B cell epitopes. Proc Natl Acad Sci USA 105(32):11311–11316. 14. Singh H, Raghava GPS (2001) ProPred: Prediction of HLA-DR binding sites. Bioinformatics 30. Hammer J, et al. (1993) Promiscuous and allele-specific anchors in HLA-DR-binding 17(12):1236–1237. peptides. Cell 74(1):197–203. 15. Southwood S, et al. (1998) Several common HLA-DR types share largely overlapping 31. Yanover C, Bradley P (2011) Large-scale characterization of peptide-MHC binding peptide binding repertoires. J Immunol 160(7):3363–3373. landscapes with structural simulations. Proc Natl Acad Sci USA 108(17):6981–6986. 16. Cantor JR, et al. (2011) Therapeutic enzyme deimmunization by combinatorial T-cell 32. Fleishman SJ, et al. (2011) RosettaScripts: A scripting language interface to the epitope removal using neutral drift. Proc Natl Acad Sci USA 108(4):1272–1277. Rosetta macromolecular modeling suite. PLoS ONE 6(6):e20161. 17. Pedelacq JD, Cabantous S, Tran T, Terwilliger TC, Waldo GS (2006) Engineering and 33. Moon JJ, et al. (2007) Naive CD4+ T cell frequency varies for different epitopes and characterization of a superfolder green fluorescent protein. Nat Biotechnol 24(1):79–88. predicts repertoire diversity and response magnitude. Immunity 27(2):203–213. 18. Brazelton TR, Blau HM (2005) Optimizing techniques for tracking transplanted stem 34. Pastan I, Beers R, Bera TK (2004) Recombinant Immunotoxins in the Treatment of cells in vivo. Stem Cells 23(9):1251–1265. Cancer. Antibody Engineering (Springer, Berlin), pp 503–518.

8582 | www.pnas.org/cgi/doi/10.1073/pnas.1321126111 King et al. Downloaded by guest on September 26, 2021